🚀 Emergent Prompting Cheat Sheet

1 · The first-prompt template

Create an app for [PURPOSE] that includes:
- [FEATURE 1]
- [FEATURE 2]
- [FEATURE 3]
- [DESIGN STYLE / vibe]

• Start with a one-sentence objective so it doesn't guess your goal.
• 3–5 features for the first build — keep scope small.
• Add one line of "vibe" for design (e.g. "clean dashboard, light theme") — don't over-specify visuals yet.

2 · Writing the initial prompt

• Describe features as USER ACTIONS, not technical tasks.
  Good: "Managers get an email when a task is marked Review and approve/reject from a button in that email."   Weak: "approval system."
• Use bullet / numbered lists, not walls of text — especially for more than 5 actions.
• List roles + permissions (admin / member / viewer) explicitly.
• List the pages/screens you need (Landing, Login, Dashboard…).
• Name integrations & data sources early (Stripe, Google login, CRM, API) — late additions force expensive rebuilds.
• Add small examples — sample inputs/outputs/messages — and state success conditions.
• Functionality first, design later. Asking for both at once gives "half-baked results on both."

3 · Iterating (after the first build)

• Let the planning Q&A run; answer briefly, then say "Start building based on what we discussed."
• One change at a time. Bundling fixes "confuses the AI." Short, targeted follow-ups.
• Use spatial language for layout: "left / right / top / bottom," "place the 3 images in one row."
• Bug protocol: test it yourself first → grab the actual error (F12 console) → say expected vs. actual → screenshot → one error per prompt, don't mix bugs with feature requests.
• If it fails, rephrase & simplify — don't repeat. 2–4 tries is normal.
• Copy your project before a major feature so you can roll back.

💰 4 · Fewest credits

• Small, focused initial scope (3–5 features) → fewer rebuilds.
• Structured lists → fewer misbuilds → fewer costly rebuild cycles.
• One change at a time → avoids expensive regressions.
• Don't over-spec design early → avoids double rebuilds (logic + UI).
• Roles / workflows / integrations up front → avoids re-scaffolding auth, routing, database.
• At ~10k records: prompt "Add indexes on the most-queried fields" → load times drop 5–10×.
• Name exact files when changing existing code (our addition) → stops credit-burning "exploration."

+ Our hard-won additions (PeptidePal)

• Repeat your guardrails + a "do NOT touch" list every prompt (summaries drop context).
• When low on credits, do single-item prompts so a mid-run cutout never leaves half-finished code.
• Verify every "files changed" summary + keep the test suite green — re-run after each change (an overlapping edit once silently broke a model).
• Don't blindly trust its code-review — it told us to break correct Python once.
• Never let it commit .env — placeholders only.